Deep learning methods for the automatic quality detection of wood logs from x-ray and shape measurements

Computed Tomography (CT) is an imaging technique used to reconstruct detailed cross-sectional images of a body by capturing a series of x-ray images with rotating sensors. CT images of logs can be acquired and processed to extract significant features that help determine the quality of the wood itself. However, CT scanners are an expensive and slow technology that may not be best suited for all use cases. This thesis presents a deep learning-based computer vision system designed to extract the same features typically derived from CT images, using only two x-ray projections and shape measurements. Specifically, the system uses Convolutional Neural Networks (CNNs) to perform three different tasks: log density extractions, pith and heartwood localization and knot detection. For the first task, a ResNet-like model is employed for the regression of different density features of the log. The second task consists of the segmentation of pith and heartwood with the use of a U-Net-like architecture. The third task leverages YOLOv5 to address the localization of knots as an object detection task. The system achieves satisfactory performance, enabling the possibility of deriving fundamental information about wood log quality with a fast and simple infrastructure, opening up the use of the sawmill process optimization to a much wider range of users.